The Uncomfortable Truth About Who Controls AI Now

In the space of two weeks this June, the US government did something it had never done so openly: it stepped into the labs of Anthropic and OpenAI and told them what they could and couldn't ship. It didn't pass a law. It reached for export controls — the machinery used for weapons — and treated a frontier AI model like a munition. Anthropic was forced to pull two of its newest models entirely. OpenAI was asked to hand out its next model one government-approved customer at a time.

So the obvious question, the one in the title, is: has Washington just taken control of AI? The short version — and the uncomfortable part — is that control hasn't disappeared, and it hasn't been captured by the government either. It has moved: off the model itself and onto compute, liability, and the people who can direct, defend with, and verify these systems. That shift should unsettle the people cheering this on and the people panicking about it in equal measure.

We'll walk through what actually happened, why the government's grip is both stronger and weaker than it looks, and — the part that matters for your career — where the real power has quietly gone. Let's get into it.

What Just Happened — the AI Regulation News Nobody Connected

Taken one at a time, the headlines looked like separate stories. Put them next to each other and a pattern jumps out.

In mid-June 2026, the US Commerce Department issued an export-control directive that forced Anthropic to pull its two newest frontier models, Fable 5 and Mythos 5, completely — not restrict them, pull them — over concerns about foreign-national access and the models' cyber-attack capabilities. Anthropic complied but said publicly the action wasn't "transparent or fair [or] grounded in technical facts." Outside experts called it arbitrary.

Days later, the same logic reached OpenAI. The Trump administration asked OpenAI to stagger the release of its next model, GPT-5.6, citing similar "Mythos-like" cyber capabilities. The detail that should stop you in your tracks: during the preview period, the government would approve access customer by customer — not "here are the rules," but signing off on who gets to use it, one at a time. Sam Altman confirmed the plan in a memo to staff and, tellingly, added that this was "not our preferred long-term model" for how releases should work.

Underneath both moves sits a policy shift. On 2 June 2026, President Trump signed Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security," which created a nominally voluntary framework for the federal government to vet powerful models before release. Three days later, on 5 June, the White House issued a separate national-security memorandum on AI. The distinction matters: the executive order set up the release-vetting frame; the memorandum signalled the national-security logic now wrapped around the whole field. Either way, it's a genuine reversal — not long ago this administration's line was that excessive regulation could kill the industry.

And none of this came out of nowhere. It's the latest step in a journey that's been building for nearly three years:

Look at the shape of it. We went from broad principles, to a deregulatory swing, to — in 2026 — the government reaching into specific products before they ship. The useful label for this is regulation by enforcement: there's no stable published rulebook, just case-by-case intervention, where each action becomes the next boundary marker.

So that's what happened. Now here's why it matters more than the headlines suggest.

The Government Just Treated AI Like a Weapon

The tool the government reached for wasn't a new AI law. It was export control — the same legal machinery used for missiles, encryption, and nuclear materials. By invoking foreign-national access and restricting who can use the models, it's effectively treating a frontier model the way it'd treat a weapons technology.

That invites the comparison everyone reaches for: is AI the new nuclear weapon? It's a useful analogy — but only if we're honest about where it holds and where it breaks, because the differences are the whole point.

With a nuclear weapon, the hard part is the raw material. Enriched uranium is scarce, slow to produce, and detectable. You don't need every hospital and city armed to defend against a bomb. Frontier AI is almost the opposite. The inputs are abundant, and the finished weights behave like a file — once released, they can be copied, mirrored, and redistributed far faster than any regulator can recall them. Running them well is another matter, and that distinction is exactly where the real control story starts.

I want to be careful here, because this is where a lot of the commentary overreaches, and I don't want to. It's tempting to say "so the government is wasting its time." That's not right. There's a real difference between the frontier capability the government is trying to gate today and the commoditised capability already loose in the world. Gating the cutting edge for a window of time is a different thing from un-inventing what's already shipped. Hold that thought — it comes back.

Why the Off-Switch Doesn't Work the Way They Think

Can the government actually contain AI? Yes and no — and the "no" is subtler than the cheerleaders for either side admit.

Start with what can't be contained: the model weights themselves. Once a capable model is released openly, no directive recalls it. It's mirrored on hard drives in a hundred countries. And the most-used models on neutral platforms aren't even American — Chinese open-weight models like DeepSeek, Qwen and GLM keep shipping freely. By early 2026, reporting put them at roughly 61% of the tokens flowing through OpenRouter's top-ten models — not the whole platform, but still enough to make the point. So when the US gates a closed American model, the customer turned away often has a free alternative the same afternoon.

But the weights were never the real chokepoint. Training and running a frontier model needs enormous concentrated compute: advanced chips, specialised memory, vast data-centre power. And that is scarce, detectable, and gateable — it's exactly why chip export controls already exist.

Here's the paradox that creates, and it's worth saying plainly: gating compute doesn't stop proliferation, because the highly capable foreign open-weight models are already out and don't depend on US chips to run. So using the one control mechanism that actually works mostly kneecaps US competitiveness while accelerating the shift to ungoverned foreign models — the exact problem it's meant to solve.

So the honest verdict isn't "the government failed." It's that its control over AI is partial, costly, and competition-distorting — not absolute, but not nothing. Which is a far less comfortable conclusion than either camp wants. The off-switch is real; it just only works on part of the machine.

So What Are We Actually Trying to Stop? (And Is AI Dangerous?)

Underneath all this sits a question worth asking plainly: is AI dangerous enough to justify any of it? The honest answer is that "AI risk" isn't one thing — and this is where I have to push back on an argument I've made myself.

In cybersecurity — my field — there's a strong case that defenders need frontier-grade capability in their own hands. The same model that finds software vulnerabilities for an attacker finds and patches them for a defender. We saw this at DARPA's AI Cyber Challenge: in the 2025 final, autonomous systems identified 86% of the contest's synthetic vulnerabilities, discovered 18 real, previously-unknown vulnerabilities in live open-source software, and submitted patches in an average of 45 minutes. In cyber, locking the best capability away from defenders arguably makes us less safe.

But that's a cyber argument, and it doesn't transfer cleanly to everything else. The key isn't that cyber is the only domain where openness ever helps — it's that cyber has the strongest symmetry: defenders genuinely need the same capability attackers use. Other domains don't share that symmetry, and some don't have it at all. Take biology: there's no version of "bio-defence needs an offensive bioweapon-design model in every lab." More capability in more hands, for that specific risk, is plausibly net-negative. Here's the spread:

So the symmetry that makes openness look obviously right in cyber gets weaker as you move down the list, and disappears entirely for bio. That's not an argument against my field's position — it's a reason to be precise about it. Cyber is the one place where gating clearly backfires, which is exactly why bio is treated differently.

And there's a serious counter-argument I won't dodge, because the safety people are right about it: even if a capability eventually becomes free, gating the cutting edge for a year or two buys a containment window. "It'll leak eventually" doesn't answer "it's contained right now, while we build defences." Safety often lives in that window. The fair criticism of the current approach isn't that buying time is pointless — it's whether case-by-case, customer-by-customer enforcement is worth its cost in distorted competition and a shove toward ungoverned foreign models.

Who Controls AI Now

So — who actually controls AI?

The honest answer is that control didn't vanish and it didn't get captured by Washington. It fractured into a stack. At the bottom, whoever owns the compute — chipmakers, governments, the few firms with the data centres. At the top, whoever carries the liability and insurance for deploying these systems — increasingly the real gate on what gets shipped commercially. And in the middle, where the actual work happens, control belongs to whoever can direct, defend with, and verify the models nobody can fully lock down or fully trust.

That middle layer is the one most people aren't looking at, and it's where my own thinking has shifted. We call open models "free," and at the level of the file, they are. But here's the part the industry keeps glossing over.

For an enterprise, that's not a free lunch; it's a supply-chain exposure of a kind security people will recognise instantly. A model's behaviour can be conditioned in ways no firewall will catch, because there's no network call to inspect — the risk is baked into the artefact. So the freedom is real at the weights layer, but trust doesn't become free — it re-concentrates at the provenance layer. And that's precisely where new, durable value is forming: the ability to verify what a model is and whether you can run it. The genie is out of the bottle on capability. It is very much not out on trust.

What This Means for Your Career and Your Skills

Here's why this matters beyond making sense of the news. If control over AI has settled into that stack — compute at the bottom, liability at the top, verification in the middle — then the layer you can actually own is the middle one. So the career move is the same as the conclusion: become one of the people who can wield, defend, and verify these systems.

This is the opposite of the doom narrative. The fear is "AI will take my job." The more accurate version, the one I keep coming back to, is that AI won't take your job — someone who understands AI will, and increasingly they'll do the work of several people. That's not a threat if you're the one doing the understanding.

For our field specifically, this regulatory shift is almost a gift. It's putting a national-security spotlight on exactly the skills cybersecurity people are built for: defending systems with AI, and the trust-and-provenance discipline of asking "can I actually rely on this model I can't see inside?" Those aren't commodities a free download replaces. They're professions.

One concrete first step, and make it a security one: take a piece of AI-generated output you'd normally accept — a suspicious script you've asked it to triage, a CVE summary, a detection rule, a chunk of code — and treat the model as if it's compromised or hallucinating. Build a verification loop that proves the output is safe to use before you'd sign your name to it. The future of this work isn't generating the output; it's auditing it. (I went deep on why the value flows to the skilled user, not the model owner, in Everyone's Watching the Wrong AI Story — this regulatory turn is that thesis playing out in real time.)

The Bottom Line

Has the government taken control of AI? The uncomfortable truth is: it took more control than the optimists thought possible — and less than it would need to actually contain the technology. It reached for the off-switch and found the switch only works on part of the machine. Control didn't vanish; it fractured — to compute, to liability, and to the people who can direct, defend, and verify these systems.

So don't build your career around predicting the next executive order. Build it around the layer every regime will still need: the people who can use AI under pressure, verify what it did, and say honestly whether it can be trusted. Whatever happens to the share prices and the next EO, that's where the power is moving — and that's the bet I'd make.

AI Regulation FAQ: Your Questions Answered

About the Author

Nathan House, Founder & CEO of StationX

Nathan House has 30 years of hands-on cybersecurity experience and is Cambridge-educated, holding CISSP, CISA, CISM, OSCP, CEH, and SABSA. He founded StationX in 1999 — one of the UK’s first cybersecurity companies — and has secured £71 billion in UK mobile banking transactions and the London 2012 Olympics, advising clients including Microsoft, Cisco, BP, Vodafone, and VISA. He authored the world’s most popular cybersecurity course — a #1 Udemy bestseller taken by over 500,000 students — and was named Cyber Security Educator of the Year 2020, AI Security Educator of the Year, and a UK Top 25 Security Influencer 2025. A DEF CON speaker and featured expert on CNN, Fox News, NBC, and the BBC, Nathan leads StationX’s training of more than half a million students worldwide.